LLM Rate Limits in 2026: How API Caps, Usage Tiers, and Cache Costs Work

The Units: RPM, TPM, RPD, and Concurrency

Four numbers describe almost every API rate limit. Requests per minute (RPM) caps how many separate API calls you can make. Tokens per minute (TPM) caps the total tokens (input plus output) flowing through in a minute. Requests per day (RPD) is a coarser daily ceiling that mostly bites free and low tiers. Concurrency caps how many requests can be in flight at once, which matters for batch and parallel workloads.

TPM is usually the one that surprises people. A single long-context request can carry hundreds of thousands of input tokens, so a handful of big calls can blow through a TPM ceiling even though your RPM is nowhere near the cap. When you design retry and queueing logic, budget against TPM first.

OpenAI Rate Limits (2026)

OpenAI organizes accounts into usage tiers, Tier 1 through Tier 5. You move up by spending cumulatively and waiting out a minimum time since your first successful payment. Each tier unlocks higher RPM and TPM per model, with reasoning models (the o-series) generally getting tighter limits than the GPT-4.1 family because they consume more compute per call. The free tier is heavily restricted, often a few requests per minute on the reasoning models, which is fine for testing and useless for production.

Tier thresholds and per-model limits shift through the year. Check your live limits in the OpenAI dashboard under Limits, and read the official rate-limits guide for current numbers.

Anthropic Claude Rate Limits and Cache Write Cost (2026)

Anthropic uses a comparable tiered model for Claude. New accounts start low; cumulative spend and account age unlock higher RPM and TPM. Claude also distinguishes input TPM from output TPM in its limit accounting, which matters because output is the slower, scarcer resource. As with OpenAI, the way to a bigger limit is to keep paying and, for large workloads, to ask the sales team directly.

The cache write cost question comes up constantly, so here is the shape of it. Anthropic prompt caching lets you store a stable prompt prefix (a long system prompt, a document, few-shot examples) and reuse it across calls. Writing that prefix into the cache costs a premium over the base input rate, commonly around 25 percent more for a standard cache lifetime, with a higher premium for longer-lived caches. Reading from the cache on later requests is the payoff: cache reads run roughly 90 percent cheaper than base input. So you pay a little extra once to write, then save heavily on every repeat. For an Opus 4.6 prompt where base input is $5 per million tokens, the cache write is billed above that and cache reads land near $0.50 per million. Confirm the exact multipliers and cache durations on Anthropic's pricing page, since they have changed more than once.

Google Gemini Rate Limits (2026)

Gemini's free tier is the most generous of the three for prototyping, offering Gemini 2.5 Flash under modest caps (on the order of low-tens of requests per minute and around a million tokens per minute, with a daily request ceiling). Paid tiers raise these substantially. Google also adjusts limits by model, with the Pro tier capped tighter than Flash. Because Google prices and limits long-context requests differently above a context threshold, a single very large request can count against your quota more heavily than its request count suggests.

Rate Limits vs Usage Limits vs Image Limits

These three get conflated. Rate limits are short-window throughput caps (RPM, TPM). Usage limits are total-consumption caps over a billing period: a monthly spend ceiling you set yourself to avoid runaway bills, plus any daily token cap your tier imposes. Image generation limits are a consumer-app concept; ChatGPT Plus and the Gemini app cap how many images you can generate per day, which is unrelated to the API token quota a developer sees. If your question is "why did my chatbot stop at 40 images," that is an app image limit, not an API rate limit.

How to Raise Your Limits Without Waiting

First, prepay credit and keep invoices current; tier increases are largely automatic once you cross spend and time thresholds. Second, batch non-urgent work through the provider's batch endpoint, which uses a separate, more generous quota and costs half. Third, cache aggressively so the same workload consumes fewer fresh tokens. Fourth, implement exponential backoff with jitter on 429s so a brief spike does not cascade into a thundering-herd retry storm. Fifth, for genuinely large or spiky production traffic, email the provider's sales team with your projected TPM; custom limits exist but you have to ask. None of these is a magic instant upgrade, but together they buy real headroom.